Shadow AI: detect and measure AI services
leak0 detects AI services in browsers and AI desktop apps and reports their use. It never blocks: blocks come from your browser policies or your MDM, leak0 checks them and provides templates.
Requirements
| Item | Requirement |
|---|---|
| Plan | Enterprise with a dedicated instance, cloud devices only. PoisonZero enables the feature for your account. |
| Operating system | Detection on macOS and Windows. Linux: browser policy check and installed AI apps. |
| macOS permissions | Accessibility for the window title, Automation per browser for the address of the active tab. The device asks once per permission. |
What is detected
| Source | Detection |
|---|---|
| Browser | Every 2 seconds: the foreground window title names the service and the address of the active tab belongs to it. |
| Running desktop app | Every 60 seconds by bundle ID (macOS) or program name (Windows). Default list: ChatGPT, Claude, Perplexity, Poe, on Windows also Microsoft Copilot. |
| Installed AI app | Once a day, even when it is not running: macOS in /Applications, Windows in programs and Store apps, Linux in .desktop files. |
| Your own AI service | By its address, optionally also as an app by bundle ID, program name or Store package family name. |
| Browser | Detection on macOS | Detection on Windows |
|---|---|---|
| Chrome, Edge, Brave | Yes, private windows skipped | Only after you allow private windows |
| Vivaldi, Arc | Yes, private windows skipped | Off, private windows cannot be told apart |
| Safari | Only after you allow private windows | Not available |
| Firefox | Not available, no address of the active tab | Only after you allow private windows |
| Opera, ChatGPT Atlas | Not available | Not available |
Turning it on
- In the cloud profile (Settings → Profiles), turn on Shadow AI and browser check.
- In the leak0 settings of the console, enter the notice text and the link to the allowed tool.
- Save. Devices take both at their next sync.
For an AI service that is not allowed, the device shows the notice, at most once a day per service. With Request approval the person asks for approval.
Anonymous mode and compliance mode
| Mode | What you see |
|---|---|
| Anonymous (default) | Rounded device counts per AI service, group and week, approval requests and browser findings per device. Never: who, which device, when, addresses or input. |
| Compliance | Adds incidents per device: secrets on their way to an AI service or agent that is not allowed, and the daily use of an AI service that is not allowed. Only this mode sends to your SIEM. |
How counting works
- A device counts once per service and week. A week appears 14 days after it ends.
- Below 3 devices nothing is shown, 3 to 9 shows as “< 10”, from 10 it is rounded to 5. Shares (Count | Percent) appear rounded or as a range.
- A group appears from its minimum size (default 10, at least 5). If its device count changes by 1 or 2, it is hidden; if only one group is hidden, the next smaller one is hidden too.
- Not all browsers covered counts devices with at least one browser that is not monitored.
In small groups, combining several values can still reveal individual use. As the controller, that is your responsibility.
Allowed AI services
AI services are not allowed by default, known agents from the PoisonZero agent list are. An approval applies to all devices or to selected profiles; on a conflict, Do not approve wins. Allowed services trigger no notice and are not blocked in newly downloaded templates. Rolling out the template is up to you.
Browser policies
Every device checks its browsers daily. The console shows a status per device and lists open gaps with the number of affected devices.
| Status | Meaning |
|---|---|
| Open gap | A browser without an enforced policy, a portable copy, an unsupported browser (Opera, unknown browsers) or no block for services that are not allowed. |
| Review | Built-in AI on, a required policy missing, Google accounts not restricted or a browser only partly supported (Vivaldi, Arc, ChatGPT Atlas). |
| OK | The check found no gap. |
| Not checked | No report, or the last check is older than 7 days. |
The templates for Chrome, Edge, Brave, Firefox and Safari block AI services that are not allowed, turn off the built-in AI of the browsers and restrict Google accounts to your domains. For a group with its own approvals, download a template of its own. Roll out to a test group first.
| Template | Format |
|---|---|
| Intune | PowerShell platform script for HKLM\SOFTWARE\Policies. |
| GPO | .reg file through a startup script, or the same values through ADMX. |
| macOS | .mobileconfig for Intune or Jamf Pro, with a content filter for Safari. |
| Linux | Shell script run as root for the policy files under /etc. |
Vivaldi and Arc have no documented policy location: set the values yourself or block the browser.
Compliance mode and SIEM
- In the leak0 settings, open Compliance mode and confirm the disclaimer.
- Within 24 hours, confirm the legal requirements. Both steps need a sign-in from the last 5 minutes.
- Set the retention: 30 to 180 days, default 90.
- An incident names device, time, agent or app, target, action and status. Never: secret value, command, path, address or prompt.
- Use counts once per device, service and UTC day and is sent at a random time, never at the moment of use.
- Secret incidents need Agentic DLP in mode On or Observe; Observe counts only with Also in the console (cloud audit) in the profile.
- If the confirmed texts change, the anonymous mode applies again. Views and deletions are logged.
Incidents go to your SIEM through the SIEM export (fields: leak0 DLP incident) or to Falcon Next-Gen SIEM.
Checking the device
poison0 doctor shows state, permissions, detection per browser, the last check and what goes to the SIEM in its leak0 DLP lines.
| Shown | Fix |
|---|---|
not licensed for this device | Check that the feature is enabled and wait for the device's next sync. |
off in the profile | Turn on Shadow AI and browser check in the profile. |
on, not running | Nobody is signed in to the desktop. Sign in and run poison0 doctor again. |
permission missing (macOS) | Allow the poison0 agent under System Settings → Privacy & Security → Accessibility and Automation. |
not available (app not signed) (macOS) | Detection needs a signed build. The browser check runs anyway. |
Detection off | Allow private windows for this browser in the leak0 settings of the console. |